# Sim-to-Real Domain Adaptation for Early Alzheimer’s Detection from Handwriting Kinematics Using Hybrid Deep Learning

## TL;DR

This paper introduces an AI system using handwriting motion data to detect early signs of Alzheimer's disease, offering a non-invasive and scalable solution.

## Contribution

A novel hybrid deep learning model with Sim-to-Real Domain Adaptation for early Alzheimer’s detection using handwriting kinematics.

## Key findings

- Classical ML models achieved moderate performance (AUC: 0.62–0.76) in detecting Alzheimer’s.
- The hybrid CNN-BiLSTM model showed high accuracy (0.91) and AUC (0.96) in early AD detection.
- Sim-to-Real Domain Adaptation improved model performance by augmenting training data with synthetic samples.

## Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by cognitive and motor decline. Early detection remains challenging, as traditional neuroimaging and neuropsychological assessments often fail to capture subtle, preclinical changes. Recent advances in digital health and artificial intelligence (AI) offer new opportunities to identify non-invasive biomarkers of cognitive impairment. In this study, we propose an AI-driven framework for early AD based on handwriting motion data captured using a sensor-integrated Smart Pen. The system employs an inertial measurement unit (MPU-9250) to record fine-grained kinematic and dynamic signals during handwriting and drawing tasks. Multiple machine learning (ML) algorithms—Logistic Regression, Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbors (kNN)—and deep learning (DL) architectures, including one-dimensional Convolutional Neural Networks (1D-CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-BiLSTM network, were systematically evaluated. To address data scarcity, we implemented a Sim-to-Real Domain Adaptation strategy, augmenting the training set with physics-based synthetic samples. Results show that classical ML models achieved moderate diagnostic performance (AUC: 0.62–0.76), while the proposed hybrid DL model demonstrated superior predictive capability (accuracy: 0.91, AUC: 0.96). These findings underscore the potential of motion-based digital biomarkers for the automated, non-invasive detection of AD. The proposed framework represents a cost-effective and clinically scalable informatics solution for digital cognitive assessment.